Editorial illustration for Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models
Liquid AI Releases LFM3: 6 Open Multimodal Models
Liquid AI has shipped LFM3, a family of six open-weight models that all trace back to a single base checkpoint, LFM3-1.7B-Base. The lineup spans text, vision, audio, combined "omni" input and a set of decision models built for speed rather than conversation, each landing in the 2-billion-parameter range. LFM3-1.7B handles text and comes post-trained for agent harnesses like Hermes Agent, Pi and OpenClaw.
LFM3-VL-2B adds video at 2 frames per second, the first LFM vision model to take moving images. LFM3-Audio-2B speaks and listens in 18 languages with tool calling built in, while LFM3-Omni-2B folds vision, audio and language together and answers back in text or speech. Two more models, d1-2B and d1-Omni-2B, are built to answer multiple questions in a single forward pass rather than generating token by token.
Every model in the family fits in roughly 1 GB of memory at 4-bit precision and ships in GGUF, MLX, ONNX, FP8 and NVFP4 formats, which Liquid AI says makes fine-tuning possible on one GPU and deployment practical on phones and laptops. How that translates into benchmark numbers is where the company's own comparisons come in.
LFM2 was tuned on 1 phone and 1 laptop. LFM3 was tuned across 14 device and backend combinations. These include a Raspberry Pi 5 and the Qualcomm Hexagon NPU.
Why this matters
LFM3's pitch is simple: one backbone, six models, all small enough to tune on a single GPU and ship to a phone. That matters more for developers and founders than the benchmark table does. Daily-Omni score of 70.1 against Qwen2.5-Omni-7B's 61.3 is a real result worth noting, since Liquid AI is beating a model more than three times the size on that task. But the video score of 61.2 and agentic score of 42.5 tell a different story, both still behind larger API-served models, which means teams building video-heavy or agent-heavy products shouldn't swap in LFM3 expecting parity.
What we'd watch next is whether that 2B-class efficiency holds up outside curated benchmarks, on messy real-world audio and multi-step agent chains where bigger models still have an edge. The audio numbers, 6.2 WER and 3.8 UTMOSv2, are respectable for a model this size, and if Liquid AI can close the video and agentic gap without ballooning parameters, on-device multimodal deployment gets a lot more practical. Right now it's a strong efficiency play with visible limits, not a drop-in replacement for frontier API models.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives NVIDIA RTX Spark Adds Native ARM64 and UE 5.8 Support for DLSS Projects
From Archives OpenAI's Ted Sanders says model names can't be reduced to a single IQ score
Common Questions Answered
What are the six models included in Liquid AI's LFM3 family?
The LFM3 family consists of six open-weight models all derived from the LFM3-1.7B-Base checkpoint: LFM3-1.7B for text, LFM3-VL-2B for video and vision, an audio model, an omni model for combined inputs, and decision models optimized for speed rather than conversation. Each model operates in the 2-billion-parameter range, making them small enough to tune on a single GPU and deploy to mobile devices.
How does LFM3's device optimization differ from LFM2?
LFM2 was tuned on only 1 phone and 1 laptop, while LFM3 was tuned across 14 different device and backend combinations including a Raspberry Pi 5 and the Qualcomm Hexagon NPU. This expanded optimization approach enables LFM3 to perform more effectively across diverse hardware platforms and edge devices.
What is LFM3-VL-2B's key capability compared to previous LFM vision models?
LFM3-VL-2B is the first LFM vision model to handle video input, processing moving images at 2 frames per second. This addition to the vision lineup allows the model to process temporal information from video content, expanding beyond static image analysis.
How does LFM3's performance on the Daily-Omni benchmark compare to larger models?
LFM3 achieves a Daily-Omni score of 70.1, which outperforms Qwen2.5-Omni-7B's score of 61.3 despite being more than three times smaller in size. However, LFM3's video score of 61.2 and agentic score of 42.5 still lag behind larger API-served models in those specific tasks.
What makes LFM3's single backbone approach significant for developers?
Having one backbone that powers six different specialized models allows developers and founders to deploy efficient, task-specific models that are small enough to tune on a single GPU and ship directly to phones and edge devices. This approach balances capability with practical deployment constraints that matter more for real-world applications than raw benchmark performance.
Further Reading
- Build specialized multimodal models with LFM3 - Hugging Face
- Liquid AI opens six-model LFM3 family at 256K context - AI Weekly
- LFM3 登場：特化型インテリジェンスのためのあらゆるモダリティ - Liquid AI
- Text Models - Liquid Foundation Models - Liquid AI Documentation
